Pedestrian Detection Method Based On Anchorless Deep Learning

Pedestrian detection technology is an important research direction in the field of computer vision,which is defined as the detection and location of pedestrians in the input image or video by computer.Pedestrian detection is widely used in autonomous driving,video surveillance,robotics and other fields.Pedestrians present different scales due to their different distances from the camera,among which small-scale pedestrians account for a large proportion.There are two types of small-scale pedestrians: those who are far from the camera and those who are short.Small-scale pedestrians have low resolution,limited features that can be extracted,and are more likely to be affected by illumination and shelter,resulting in missed detection and misdetection of pedestrians.Although the pedestrian detection method for the application of deep learning has made great progress,but because of the small-scale pedestrian detection effect is still cannot meet the needs of actual industry,therefore this article in view of the existing pedestrian detection algorithm of small-scale pedestrian information describing the problem of inadequate was improved,the main research work is as follows:1.Summing up the ResNet residual network and its improvement methods based on CSP pedestrian detection algorithm.Firstly,the foundation of convolutional neural network is introduced.Secondly,it analyzes the advantages and disadvantages of the application of ResNet residual network in pedestrian detection and six improvement methods of ResNet residual network proposed in recent years,and specifically analyzes its improvement points,improvement framework and advantages and disadvantages.Finally,three typical residual-like networks,IResNet、ResNet St and Vo VNet,were selected as the backbone network of CSP algorithm and named as I-CSP,S-CSP and V-CSP algorithm respectively.Experimental comparison and analysis were conducted with other algorithms on the pedestrian detection data set of City Persons.2.A pedestrian detection algorithm based on self-calibrated convolutional network(SCCSP)is proposed.In this algorithm,the basic network of CSP algorithm is improved based on the network structure of SCNet self-calibration convolution.Specifically,firstly,SCNet selfcalibration convolutional network is used as the backbone network to expand the range of the network’s receiver field,so that the detector proposed has a strong multi-scale representation capability for pedestrians.Secondly,in the feature extraction stage,the image is input into the self-calibrated feature extraction network,and then the semantic information of the high and low layers is fused layer by layer from bottom to top.Finally,pedestrian detection is simplified to a direct center point and scale prediction task by multi-layer connection of fine multi-scale convolution features.The algorithm and the comparison method were tested on the pedestrian data sets of City Persons and Caltech respectively.The experimental results show that the algorithm presents a very good pedestrian detection performance,especially improving the detection effect of small-scale pedestrians.